A Novel Approach to Optimizing Deep Learning Models for Edge Computing Devices

Abstract

The proliferation of edge computing devices has created an urgent need for deploying deep learning models on resource-constrained hardware. Traditional deep neural networks (DNNs) are designed for powerful cloud-based GPU clusters, making their direct deployment on edge devices impractical due to limited computational resources, memory constraints, and power budgets. This paper proposes a novel multi-stage optimization framework that combines structured pruning, knowledge distillation, hardware-aware quantization, and layer fusion to significantly reduce model complexity while preserving predictive accuracy. Our framework, called EdgeOptNet, achieves a 71.5% reduction in memory footprint and a 57.3% reduction in inference latency compared to baseline models, while maintaining 96.2% of the original top-1 accuracy on ImageNet. We validate our approach on four commercial edge platforms: Raspberry Pi 4, NVIDIA Jetson Nano, Google Coral Edge TPU, and Intel Neural Compute Stick 2. Extensive experiments demonstrate that EdgeOptNet consistently outperforms existing optimization techniques across all evaluated metrics, including accuracy, latency, memory consumption, and power efficiency. We also present a comprehensive ablation study analyzing the contribution of each optimization stage to the overall performance improvement.